Spaces:
Running on Zero
Running on Zero
File size: 2,498 Bytes
4ee3e9f bca4977 4ee3e9f bca4977 4ee3e9f 73a98d6 4ee3e9f 09d4474 73a98d6 4ee3e9f bca4977 4ee3e9f bca4977 4ee3e9f bca4977 4ee3e9f bca4977 4ee3e9f 73a98d6 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 | """
AOT compilation optimization for Qwen-Image-Edit pipeline.
"""
import gc
from typing import Any
from typing import Callable
from typing import ParamSpec
from torchao.quantization import quantize_
from torchao.quantization import Float8DynamicActivationFloat8WeightConfig
import spaces
import torch
from torch.utils._pytree import tree_map
P = ParamSpec('P')
TRANSFORMER_IMAGE_SEQ_LENGTH_DIM = torch.export.Dim('image_seq_length')
TRANSFORMER_TEXT_SEQ_LENGTH_DIM = torch.export.Dim('text_seq_length')
TRANSFORMER_DYNAMIC_SHAPES = {
'hidden_states': {
1: TRANSFORMER_IMAGE_SEQ_LENGTH_DIM,
},
'encoder_hidden_states': {
1: TRANSFORMER_TEXT_SEQ_LENGTH_DIM,
},
'encoder_hidden_states_mask': {
1: TRANSFORMER_TEXT_SEQ_LENGTH_DIM,
},
'image_rotary_emb': ({
0: TRANSFORMER_IMAGE_SEQ_LENGTH_DIM,
}, {
0: TRANSFORMER_TEXT_SEQ_LENGTH_DIM,
}),
}
INDUCTOR_CONFIGS = {
'conv_1x1_as_mm': True,
'epilogue_fusion': False,
'coordinate_descent_tuning': True,
'coordinate_descent_check_all_directions': True,
'max_autotune': True,
'triton.cudagraphs': True,
}
def optimize_pipeline_(pipeline: Callable[P, Any], *args: P.args, **kwargs: P.kwargs):
@spaces.GPU(duration=1000)
def compile_transformer():
# Run warmup pass to capture transformer inputs (needs all components on GPU)
with spaces.aoti_capture(pipeline.transformer) as call:
pipeline(*args, **kwargs)
# Offload text encoder to CPU to free ~16GB during memory-intensive torch.export
# This happens AFTER warmup but BEFORE export to avoid device mismatch
text_encoder_device = next(pipeline.text_encoder.parameters()).device
pipeline.text_encoder.to('cpu')
gc.collect()
torch.cuda.empty_cache()
dynamic_shapes = tree_map(lambda t: None, call.kwargs)
dynamic_shapes |= TRANSFORMER_DYNAMIC_SHAPES
# quantize_(pipeline.transformer, Float8DynamicActivationFloat8WeightConfig())
exported = torch.export.export(
mod=pipeline.transformer,
args=call.args,
kwargs=call.kwargs,
dynamic_shapes=dynamic_shapes,
)
compiled = spaces.aoti_compile(exported, INDUCTOR_CONFIGS)
# Move text encoder back to original device
pipeline.text_encoder.to(text_encoder_device)
return compiled
spaces.aoti_apply(compile_transformer(), pipeline.transformer) |